Practical engineering posts from Ergini, a senior software & AI developer in Kosovo. RAG, human-in-the-loop patterns, AI scheduling, MVP cost, and the things vendor blogs skip.
A composite story: an n8n workflow stays green for a month while no trade invoices go out, and the audit, ledger and daily check that make green mean done.
Read post→
A composite story: a German customer bounces a PDF invoice, the old ERP can't make XRechnung, and a converter layer issues validated e-invoices before 2027.
Read post→
A composite story: one supplier invoice arrives twice and is paid twice, and the intake, supplier matching and duplicate checks that hold the copy.
Read post→
A composite story: reminders keep chasing an invoice the customer disputed in a reply nobody read, and the collections workflow that reads the thread first.
Read post→
A rejected candidate asks why, and the AI match score has no answer. The explainable CV screening that quotes evidence, with a recruiter deciding every no.
Read post→
One long-serving colleague answers every 'where is the template' question. The Teams assistant that answers from SharePoint, with sources and permissions.
Read post→
A haulier's 40-tab dispatch workbook breaks as the fleet grows. The internal tool that replaces it, with rules, roles and history, and keeps Excel exports.
Read post→
A physio practice's cancelled slots stay empty while patients wait weeks. The waitlist agent that offers each slot by message, under rules the practice sets.
Read post→
A composite story: a liability cap switched off by a definition on page two, and a contract review that checks every clause against the legal playbook.
Read post→
A composite story: a tenant break missing from a seller's lease schedule, and an abstraction that cites every term and computes critical dates in code.
Read post→
A composite story: a container held over a tariff code copied from a lookalike row, and a workflow that drafts customs declarations for a declarant to sign.
Read post→
A composite story: a mortgage file stalled on three missing documents, and a workflow that asks for exactly what is missing and checks every upload.
Read post→
A composite story: a Lovable app's first paying customer still sees the trial banner, and the audit that fixes data, keys and billing before the second.
Read post→
A composite story: a SaaS company's model bill triples in a month with no feature to blame, and the cost audit that tags every call and finds the cause.
Read post→
A composite story: a law firm finds client contracts in personal ChatGPT accounts, and the private AI gateway that gives staff a sanctioned route, not a ban.
Read post→
A composite story: a distributor's sales reps paste ERP exports into Claude, and the scoped MCP server that lets them ask SAP directly, without the keys.
Read post→
A composite story: a court order filed behind a brief in beA, and a docketing workflow where a model reads, code counts the days and a lawyer confirms.
Read post→
A composite story: a law firm loses evening inquiries and calls prospects back before the conflict check. The intake workflow that puts conflicts first.
Read post→
A composite story: a small practice chasing 230 clients for receipts every January, and a workflow that asks each client for exactly what is missing.
Read post→
A composite story: an edited payslip scored clean, an honest one flagged, and a fraud check that tests claims against bank data while people decide.
Read post→
A composite story: a solar installer buried in grid operator portals and MaStR entries, and an AI agent that fills each filing up to the submit button.
Read post→
A composite story: a small contractor nearly skips a 212-page public tender nobody has time to read, and an AI assistant builds its compliance matrix.
Read post→
A composite story: a heating and plumbing firm whose after-hours calls hit voicemail, and an AI phone agent that puts emergencies through and books the rest.
Read post→
A composite story: a road forwarder loses spot loads to slow replies, and the AI agent that reads quote emails in four languages and prices them by its rules.
Read post→
A composite story: a job shop wins only the RFQs its estimator reaches, and the workflow that reads drawings and costs parts from its own ERP history.
Read post→
A composite story: an MGA's broker submissions sit unread while brokers place risks elsewhere, and a workflow that reads and triages them for underwriters.
Read post→
A composite story: an agency's account managers lose Mondays to screenshots, and a reporting pipeline that drafts client commentary from proven causes.
Read post→
A composite story: a radiology group re-keys 300 faxed referrals a week, and a workflow reads, checks and chases them while clinicians keep clinical calls.
Read post→
A composite story: one on-call night at a property manager, and an AI agent that triages tenant WhatsApp requests so only real emergencies wake anyone.
Read post→
A composite story: an IT staffing agency pays to source contractors already in its ATS, and a workflow that finds them and drafts outreach for recruiters.
Read post→
A composite story: a night desk registers late guests by hand, and a workflow that files Alloggiati Web and SES.HOSPEDAJES reports on time.
Read post→
A composite story: an electrical subcontractor, a hindrance buried in a site diary, and an AI agent that flags VOB/B notice deadlines before they pass.
Read post→
My AI fashion studio's first renders were beautiful and wrong: stripes, logos and colors drifted. How I made it check every image against the real garment.
Read post→
My curtain visualizer first rendered fabrics nobody sells at the wrong scale. How I split the job: geometry and pattern in code, light and folds for the model.
Read post→
My screener first gave every applicant one score nobody could explain. How I rebuilt it around evidence per criterion, an override log and batched review.
Read post→
OmniAPI turns a one-sentence description into a typed, deployed function. Why generated code is untrusted code, and the contract, sandbox and tests around it.
Read post→
The first version of my open-source agent skill questioned me. The better one questions its own work: every doubt ends in a run, a test or the source.
Read post→
My scheduling agent read a revoked calendar token as a free day and proposed a slot on top of a meeting. The fix is one rule every tool-using agent needs.
Read post→
Every page on my Next.js site shipped JSON-LD, and AI crawlers saw none of it: next/script had turned it into JavaScript. How I found it, the fix, and the check.
Read post→
Google showed an old blue icon for my site. It was not a cache: Google does not list SVG as a supported favicon format, so it took the only PNG I had declared.
Read post→
Nineteen queries in my Search Console were never typed into a search box: a pasted prompt, follow-ups like 'until when?', Greek and Arabic. What they are and what I changed.
Read post→
First-party Search Console data on an automated agent that generated 225 permutations of a single query over three weeks, produced 3,572 impressions at average position 8.2, never clicked once, and stopped dead on a single day. How to recognise it, why it wrecks your reporting, and what it says about agents on the open web.
Read post→
The decision teams reach after their automation platform stops scaling. Where Zapier, Make, n8n and Pipedream genuinely end, the four failure modes that appear before the ceiling does, and how to split a stack so the boring 80 percent stays no-code and the load-bearing 20 percent gets built.
Read post→
Article 50 became enforceable on 2 August 2026. A builder's guide to what each of its four obligations requires, which fall on you as provider and which as deployer, the exemptions, and the 2 December 2026 marking deadline.
Read post→
A decision framework for picking an embedding model in 2026: API vs self-hosted, when dimensions matter, multilingual and multimodal, and how to benchmark on your own data in an afternoon.
Read post→
BGE-M3 against OpenAI text-embedding-3: multilingual quality, the hybrid dense-sparse-multivector trick BGE-M3 has and OpenAI does not, hosting cost, and when self-hosting actually pays off.
Read post→
Voyage AI embeddings against OpenAI text-embedding-3 for production RAG: retrieval quality on technical and domain corpora, the domain-specific model lineup, reranking, and total cost.
Read post→
Cohere Embed v4 against OpenAI text-embedding-3-large: multimodal and PDF embedding, Matryoshka dimensions, VPC deployment, multilingual coverage, and which one fits an enterprise stack.
Read post→
What a reranker does, why cross-encoders beat embeddings at final ranking, how much quality it actually adds, what it costs in latency, and how to add one to an existing RAG pipeline.
Read post→
The real limits of Pinecone's free Starter plan in 2026: storage, indexes, namespaces, read and write units, region restriction, and how many vectors that works out to in practice.
Read post→
The architecture behind a real AI scheduling agent: calendar and inbox parsing, the tool-calling loop, conflict resolution, failure modes, and what it actually costs to build and run.
Read post→
AI-generated on-model imagery is a deepfake under Article 50(4) and must be disclosed, regardless of commercial intent. What EU fashion and e-commerce brands must label, what must be machine-readably marked, and by when.
Read post→
How AI fashion photography works in 2026: generate photorealistic, on-model imagery from your own garments, when it beats a real shoot, and where it still falls short.
Read post→
A practical guide to AI product photography for fashion e-commerce: on-model vs ghost mannequin, PDP conversion, batch generation at catalog scale, and staying on-brand.
Read post→
What a fashion photoshoot really costs in 2026, line by line, and how AI-generated on-model imagery compares on price, speed, and quality for brands and stores.
Read post→
What the EU AI Act actually requires of startups after the Digital Omnibus: Article 50 transparency is live now, high-risk duties moved to December 2027, and here is the engineering work for each.
Read post→
How well do ChatGPT and Claude handle Albanian (shqip) in 2026? An honest look at quality, the gaps, and how to build reliable Albanian-language AI features.
Read post→
Seven AI automation workflows for real estate agencies in 2026: lead qualification, listing copy, document handling, follow-up, and more, with build notes.
Read post→
What an AI receptionist actually does in 2026, the build-vs-buy math, and how to set one up that books appointments and answers calls without annoying callers.
Read post→
Claude AI is not on claude.ai for Kosovo as of 2026. How to access Claude from Kosovo, the Albanian-language reality, and building on the Claude API.
Read post→
Honest 2026 guide to AI chatbots for websites: 8 platforms compared, build-your-own costs, knowledge-base sync, and hallucination control.
Read post→
Real MVP costs in 2026 from a developer who ships them solo. Line items, AI-MVP add-ons, where founders waste money, and a transparent rate card.
Read post→
Step-by-step RAG architecture tutorial with TypeScript code, retrieval evaluation, and the five failure modes I hit shipping production RAG.
Read post→
Nine AI scheduling assistants tested across 47 real meetings by the developer who shipped Caldra AI. Pricing, real capabilities, build vs buy, and what the marketing pages omit.
Read post→
Real human-in-the-loop AI patterns from production: approval queues, confidence-based routing, post-hoc audit, and active learning. With code and case studies.
Read post→
End-to-end AI voice agent on a real phone number. Twilio + ElevenLabs + GPT, with interruption handling, sub-800ms latency, and warm transfer.
Read post→
A senior engineer's take on Pinecone vs Qdrant vs Weaviate vs pgvector. Benchmarks, total cost, and the crossover point where self-hosting wins.
Read post→
Build an AI meeting transcription and summary tool with Whisper, speaker diarization, and structured output. Self-hosted, private, customizable.
Read post→
Kosovo offers CET-aligned, EU-adjacent, English-speaking senior dev talent at 60% off Western rates. The practical guide to making it work.
Read post→
A senior engineer's take on LangChain vs Vercel AI SDK for TypeScript AI apps. Architecture, bundle size, streaming, and when to use both.
Read post→
A senior AI dev's hiring playbook for founders. What to look for, how to interview, how to test for real LLM experience vs prompt-fluency theater.
Read post→
Which automation platform wins for AI workflows in 2026? n8n's AI nodes vs Zapier's polish vs Make's visual canvas - broken down by use case.
Read post→
Next.js, Supabase, Vercel, Stripe, Clerk, Resend. The exact MVP stack I have shipped 6+ products on, with the boilerplate decisions made for you.
Read post→
What it really costs to build an AI chatbot. Tiers from $5K demo to $200K enterprise, with what each tier gets you and where money is wasted.
Read post→
Build a private AI email agent for Gmail or Outlook that triages, labels, drafts replies, and respects your tone - without sending without approval.
Read post→
What OpenAI's API actually costs at scale. GPT-5, GPT-5-mini, caching, Batch API, and the cost-control patterns that cut my client bills 60%.
Read post→
pgvector handles 50M vectors on a normal Postgres box. Pinecone earns its price past 100M. Here is the honest crossover from production benchmarks.
Read post→
End-to-end guide to building a production AI support bot with RAG, escalation logic, evals, and Intercom or Zendesk integration. Code included.
Read post→
Claude Code is a terminal agent. Cursor is an AI-native IDE. I use both - here is the task-by-task breakdown of when each one earns its keep.
Read post→
A working dev's comparison of Claude Opus 4.7 vs GPT-5 on coding, context, pricing, and API reliability - based on daily shipped work.
Read post→
When to use RAG, when to fine-tune, when to do both. Real cost data, accuracy numbers, and a decision tree for picking the right approach.
Read post→
A signal-based AI lead-gen tool - intent detection, enrichment, scoring, and outbound copy - built from scratch with the Vercel AI SDK.
Read post→
Agentic RAG turns retrieval into a reasoning loop. Here is how to design, evaluate, and ship one without melting your latency budget or your token bill.
Read post→
Production guide to OpenAI structured outputs, verified August 2026 against the Responses API: strict mode, schema design, refusals, streaming, and how OpenAI, Anthropic and Gemini compare.
Read post→
Defense-in-depth strategies for prompt injection. Channel separation, output filtering, tool scoping, and the OWASP LLM Top 10 in plain English.
Read post→
Build an LLM-powered document extraction pipeline for invoices, contracts, and forms. Layout-aware OCR, schema design, validation, human review.
Read post→
Comparing the LLM eval frameworks engineers actually run in CI. Covers metrics, latency, cost, and which one to pick for RAG, agents, or chat.
Read post→
Side-by-side review of the top LLM observability platforms. Cost, integration time, framework support, and which to pick for agents vs RAG vs chat.
Read post→
MCP server tutorial in TypeScript with tools, resources, OAuth, and remote SSE transport. Wire it into Claude Code, Cursor, and Claude Desktop.
Read post→
Build a bias-aware AI resume screener: structured extraction, rubric scoring, explanations, and the audit log that keeps recruiters confident.
Read post→
Build content moderation with the free OpenAI Moderation API, then layer custom classifiers and human review for edge cases.
Read post→
How to design, name, scope, and document tools for reliable LLM tool calling. Parallel calls, error handling, prompt-injection-safe tool design.
Read post→
Workflows are predictable and cheap. Agents are flexible and expensive. Here is how to tell which one your problem actually needs, with examples.
Read post→
Supabase ships pgvector and Edge Functions; Firebase ships Genkit and Gemini integration. Here is the practical pick for AI-native apps.
Read post→
Compares the leading embedding models for RAG. MTEB scores, context length, multilingual support, and price per million tokens.
Read post→
AI dev rates broken down by region, seniority, and engagement model. Includes the hidden costs most cost guides ignore.
Read post→
Serbia, Albania, Kosovo, North Macedonia, Bosnia: rates, talent depth, English proficiency, and the practical playbook for hiring well.
Read post→
The agent patterns that actually ship: reflection, planner-executor, ReAct, multi-agent handoff, and when to use a workflow instead.
Read post→
The architecture decisions every AI SaaS converges on: multi-tenant data, per-user limits, model routing, eval pipelines, and BYO-key support.
Read post→
Embedding, vector DB, retrieval, generation. The full per-query economics of RAG at three scales - with a downloadable calculator.
Read post→
The 30-item checklist I run with every AI MVP client before writing a line of code. Eval strategy, cost guardrails, fallback model, data policy.
Read post→
A ground-level tour of Kosovo's tech ecosystem - hubs, talent pools, salary bands, key companies, and how Western founders should engage.
Read post→
The Eastern European AI talent map: Poland, Ukraine, Romania, Kosovo, Serbia. Where senior LLM engineers live and what they cost in 2026.
Read post→
Build AI sales automation that scores intent, drafts context-aware outreach, and routes to humans - not another mass-email blaster.
Read post→
A clean 2026 tutorial for streaming OpenAI responses in Next.js App Router using Server Actions and the Vercel AI SDK. With tool calls.
Read post→
How to design, ship, and debug tool calls with the Vercel AI SDK. Includes parallel tools, streaming UI, error handling, and Zod validation.
Read post→
Build an internal AI knowledge assistant from Notion, Slack, Drive, and Linear. Permissions-aware, freshness-aware, citation-first.
Read post→
When Bubble or Webflow beats a custom build, and when it doesn't. Real cost, speed, and ceiling tradeoffs for early-stage founders.
Read post→
When a freelance AI developer beats an agency, and when it doesn't. Cost, speed, quality, risk - broken down for founders.
Read post→
A pragmatic guide to building or installing an AI code reviewer for your PRs. CodeRabbit, PR-Agent, or your own GitHub Action - all compared.
Read post→